# Re: Singularity Arrival Estimate

From: Eugen Leitl (eugen@leitl.org)
Date: Thu May 16 2002 - 06:04:33 MDT

On Tue, 14 May 2002, Dani Eder wrote:

> The first is based on the idea that progress
> is proportional to the total thinking power
> on the Planet.

Progress measured in what?
Total thinking power measured in what?

Can you cite any data backing up this idea of yours?

> We start by assuming the human brain
> is worth 10^17 bits/sec of processing power
> plus or minus a factor of 100. Annual

That's an ad hoc number. No one really knows.

> cpu production is equivalent to 40 brains
> currently on that basis. When it gets to

Naked CPUs don't compute, computers compute. Computers are not all-purpose
machines as far as neuronal processing is concerned, so you would to have
to run a benchmark. Notice that you're assuming embarassingly parallel
code here.

> a 10^7 brain-equivalents a year, the thinking
> power on the planet starts to curve upward
> rapidly. On a one year/doubling for computer
> power, that's 18 years +- 7, or 2020AD+-7 years.

There's no abstract computer "power". You have to have to run your
specific benchmark. It's not obvious that we'll have all-purpose molecular
circuitry in ~two decades, allowing a smooth transition from semiconductor
photolitho to self-assembled molecular circuitry.

> The second is based on the idea that progress
> is proportional to the smartest thinking AI.

Measured in what?

AI is the archetypal thresholded development. All the hardware in the
world under your fingertips is worth jack as long as you don't know how to
use it.

> The smartest AI is assumed to be a combination
> of hardware speed and time spent 'learning',
> building it's internal information structures.
> For example, a one brain human takes ~20 years
> to accumulate enough information to be a useful
> thinker. The singularity point is then 2 years
> after a ~10 brain machine becomes available for
> AI research, assuming a 1 year Moore's Law.
> Taking the 100th machine on the TOP500 supercomputer
> list as a proxy for 'available for AI research',
> That projects a date of 2025 for the singularity
> +-7 years.
>
> Averaging the above results and knocking off 1/3
> for continuing reduction in the doubling time
> for computer power, I would make a personal
> projection of 14 years +-5 years.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:38 MDT